Is open-source AI safer or riskier than closed/proprietary AI?

There's no single, settled answer — this is one of the most actively contested questions in AI policy, and the "safer" vs "riskier" verdict depends heavily on what kind of risk you're talking about and who you ask. Here's a breakdown of the strongest arguments on each side, based on current research and debate.

The case that open AI is riskier

1. Safeguards can be stripped off. Closed labs put extensive guardrails (content filters, refusal training, usage policies) on top of their models, but open-weight models pose additional risks because their safeguards are easier to remove, since safeguards are much easier to remove for open-weight models even though both open-weight and closed models can have safeguards to refuse harmful user requests. Malicious actors can fine-tune a model to optimise its performance for harmful applications, remove parts of the code designed to prevent harmful uses, or undo previous safety fine-tuning.

2. No recall mechanism. With a hosted/closed model, the vendor retains control. Open-weight models can be spread quickly and irreversibly — if a closed-weight model is found to pose hazards, risk-conscious developers can add patches or pull the model from distribution, as happened with OpenAI's April 2025 update of GPT-4o after external evaluation identified excessive sycophancy, prompting a revert. Once weights are released publicly, there's no equivalent "undo" button.

3. Larger attack surface, less mature defenses. Frontier closed-weight developers rely on training interventions, content classifiers, and enforcement of acceptable-use policies, but none of these provide reliable assurances for open-weight models, which can be used, tampered with, and distributed without limitations — so the attack surface for open-weight models is larger, and the toolkit for defending them is less well studied.

4. Governability, not raw capability, is the real gap. One recent security analysis frames it well: open-weight models are not categorically more dangerous than closed models — they are less governable. Closed models give you enforcement you do not own; open weights give you sovereignty you must now enforce yourself.

5. Geopolitical/national-security dimension. Open source AI became part of U.S.-China geopolitical competition after the Chinese firm DeepSeek released its open source R1 generative AI model in January 2025, raising data privacy, censorship, and national security concerns, and U.S. regulators have focused on potential national security implications of open source AI, particularly if controlled information is inadvertently released.

The case that open AI is safer (or at least not inherently riskier)

1. Transparency enables independent scrutiny. Open source enables widespread community oversight as models and datasets can be interrogated, scrutinized, and evaluated by anyone, without needing to seek approval from a central decision-maker. This "many eyes" effect is the classic open-source security argument applied to AI.

2. Reduces dangerous concentration of power. With open source, any developer can hypothetically leverage the investments of larger companies, governments, and research labs, democratizing AI development and deployment by enabling people to examine, reuse, and build on top of these powerful systems — a counterweight to a handful of corporations controlling increasingly powerful systems unaccountably.

3. Even frontier labs argue closed isn't automatically safer. Cisco, Google, Microsoft, and OpenAI have all argued that abuse-risk isn't a reason to avoid open-weight models; closed models can be abused in equally hard-to-detect ways, and openness lets independent researchers scrutinize them.

4. "Open" is not monolithic, and most "open" models aren't fully transparent anyway. A useful distinction: closed/API models never release weights and run only on vendor servers, open-weight models publish weights but keep the training recipe proprietary, while true open-source (strict) requires weights plus architecture plus training code plus enough data to reproduce from scratch — with something like OLMo being a rare example. Most "open-source" AI today is actually just open-weight, so the transparency benefits are often overstated.

Where serious researchers land: it's unresolved and risk-type-dependent

Experts largely agree the picture is mixed and empirically uncertain. The potential security hazards linked with making models open-source continue to be the crux of debates among AI researchers and policymakers, the offence-defence balance remains controversial, and there is still debate over whether open-source models will increase model security or increase the risk of abuse. Similarly, a technical risk-management paper notes it is unclear how much counterfactual risk open-weight models will pose compared to closed-weight models, and cautions researchers to be mindful of both "openness washing" and "safety washing," remaining open to evidence both for and against the possibility that some models pose large risks if deployed openly.

Bottom line

So the honest answer is: neither is categorically "safer" — they shift where the risk and responsibility sit, and the right choice depends on which specific harms you're trying to prevent, who the threat actors are, and how much governance capacity you have to manage a model running outside anyone's central control.